Crawler vehicle running state prediction model training method, prediction method and product

Abstract

The application discloses a tracked vehicle driving state prediction model training method, a prediction method and products, relates to the intelligent driving technical field of tracked vehicles, and first acquires measurement vector data and measurable control vector data of a tracked vehicle; inputs the measurable control vector data into a trained transmission-track model to obtain actual control vector data; acquires state vector data of the tracked vehicle at a previous moment, sequentially calculates prior estimation of the state vector data and the measurement vector data to obtain prior state estimation and prior measurement estimation; converts the prior state estimation, the prior measurement estimation and the actual control vector data into a difference sequence; trains a recurrent neural network by using the difference sequence to obtain a tracked vehicle driving state prediction model. The application combines a neural network with good nonlinear fitting characteristics into a Kalman filtering algorithm, realizes effective state estimation of a strong nonlinear system, integrates a Kalman filtering state estimation principle, and reduces to-be-learned parameters of the neural network.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent driving of tracked vehicles, in particular to a tracked vehicle driving state prediction model training method, a prediction method and products. BACKGROUND

[0002] With the development of sensing and control technologies, tracked vehicles, as an important special vehicle, are developing towards intelligence. Due to driving on complex and changeable ground, the driving state of tracked vehicles changes rapidly, and dangerous working conditions such as turning side slip and even rollover may occur. Therefore, in order to achieve high-precision and safe driving control, it is necessary to obtain a high-reliability tracked vehicle driving state. Although GNSS and IMU sensors can directly measure state variables such as position, yaw angular velocity and acceleration, due to the fact that tracked vehicles mostly drive in harsh off-road environments with sand, rain and bumps, sensor noise cannot be ignored, so it is necessary to use a state estimation method to obtain accurate and reliable driving states.

[0003] At present, state estimation methods based on mechanism models, such as Kalman filter and its variants, require the establishment of kinematic or dynamic models of tracked vehicles before combining sensor information to estimate the state, which is suitable for linear or weakly nonlinear systems that can be modeled. However, when tracked vehicles are driving, there is a dynamic coupling between each direction, and the state variables of each direction are obviously coupled, and the motion is strongly nonlinear. The tracked force model usually uses integration of time and position based on the modeling of ground shear displacement, which is strongly nonlinear. At the same time, due to the dramatic changes and random disturbances of the ground conditions on which tracked vehicles are driving, the mismatch between the nominal model of the vehicle and the actual controlled object occurs. Strong nonlinearity and model mismatch result in poor performance of state estimation methods based on mechanism models. Data-driven state estimation methods do not require the establishment of complex high-precision models, but learn the relationship between system inputs and outputs through data, which has good applicability, but lacks interpretability, and the black-box nature leads to a large number of parameters to be learned and long training time. SUMMARY

[0004] The purpose of the present application is to provide a tracked vehicle driving state prediction model training method, a prediction method and products, which can effectively estimate the state of a strongly nonlinear system, and has fewer parameters to be learned by a neural network, thereby improving the training efficiency.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] In a first aspect, the present application provides a tracked vehicle driving state prediction model training method, which comprises:

[0007] Obtaining measurement vector data and measurable control vector data of the tracked vehicle; the measurement vector data includes longitudinal, lateral acceleration and yaw rate of the tracked vehicle; the measurable control vector data includes left and right drive motor torque, left and right drive wheel speed and transmission gear position;

[0008] Inputting the measurable control vector data into the trained drive-track model to obtain actual control vector data; the actual control vector data includes lateral and longitudinal force of the ground on the track, steering drive torque and steering resistance torque;

[0009] Obtaining state vector data of the tracked vehicle at the previous time; the state vector data includes predicted values of longitudinal, lateral speed and yaw rate of the tracked vehicle;

[0010] Calculating the priori estimation of the state vector data at the current time by using the Kalman recursive framework to obtain the priori state estimation;

[0011] Calculating the priori estimation of the measurement vector data by using the Kalman recursive framework to obtain the priori measurement estimation;

[0012] Converting the priori state estimation, the priori measurement estimation and the actual control vector data into a difference sequence;

[0013] Training a recurrent neural network by using the difference sequence to obtain a tracked vehicle driving state prediction model; the output of the tracked vehicle driving state estimator is state vector prediction data; the state vector prediction data is the real-time longitudinal, lateral speed and yaw rate of the tracked vehicle predicted.

[0014] Optionally, the training method of the drive-track model specifically comprises:

[0015] Obtaining drive data of the tracked vehicle; the drive data includes left and right drive motor torque training data, left and right drive wheel speed training data and transmission gear position training data;

[0016] Performing force analysis on the tracked vehicle to obtain vehicle force data; the vehicle force data includes lateral and longitudinal force training data of the ground on the track of the tracked vehicle, steering drive torque training data and steering resistance torque training data;

[0017] Training a neural network by taking the drive data as input and the vehicle force data as output to obtain a drive-track model.

[0018] Optionally, the neural network is a long short-term memory neural network.

[0019] Optionally, the loss function of the recurrent neural network is:

[0020] wherein, represents a loss function, Δy k represents a measurement update difference.

[0021] Optionally, the recurrent neural network comprises three independent LSTM units and a plurality of FC layers.

[0022] The three independent LSTM units are respectively used for learning a process noise covariance matrix, a state covariance matrix and a measurement covariance matrix.

[0023] Optionally, the Kalman recursion framework is a Kalman recursion framework based on a mechanism model.

[0024] The processing procedure of the Kalman recursion framework comprises:

[0025] A tracked vehicle dynamics model is established:

[0026]

[0027] wherein, m is a vehicle mass, I z is a steering inertia of the vehicle, B is a track center distance, v x is a longitudinal velocity, v y is a lateral velocity, ω is a steering angular velocity, i=1, 2 respectively represent left and right tracks, F xi is a longitudinal driving force of the ground acting on the two tracks, F yi is a lateral force of the ground acting on the two tracks, T di is a steering driving torque of the two tracks, T ri is a steering resisting torque of the two tracks, F N1 ,F N2 is a ground pressure on the left and right tracks, g is a gravity acceleration, f r is a ground resistance coefficient.

[0028] A model state variable is x=[v x ,v y ,ω] T , a control quantity is u=[F x1 ,F x2 ,F y1 ,F y2 ,T r1 ,T r2 ,T d1 ,T d2 ] T , and a measurement quantity is y=[a x ,a y ,ω] T wherein a x ,a yare the longitudinal and lateral acceleration of the vehicle respectively, and ω is the yaw rate;

[0029] The state transition function is:

[0030]

[0031] According to the tracked vehicle dynamics model, the measurement function is obtained as:

[0032]

[0033] In a second aspect, the application provides a tracked vehicle driving state prediction method, the tracked vehicle driving state prediction method comprising:

[0034] Obtaining state vector data of a previous moment of a tracked vehicle to be predicted, real-time measurement vector data and real-time measurable control vector data;

[0035] Inputting the state vector data of the previous moment, the real-time measurement vector data and the real-time measurable control vector data into a tracked vehicle driving state prediction model to obtain real-time longitudinal, lateral velocity and yaw rate of the tracked vehicle to be predicted; the tracked vehicle driving state prediction model is trained by the tracked vehicle driving state prediction model training method of any one of the above.

[0036] In a third aspect, the application provides a computer device, comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer program to implement the tracked vehicle driving state prediction model training method or the tracked vehicle driving state prediction method of any one of the above.

[0037] In a fourth aspect, the application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the tracked vehicle driving state prediction model training method or the tracked vehicle driving state prediction method of any one of the above.

[0038] In a fifth aspect, the application provides a computer program product comprising a computer program, wherein the computer program is executed by a processor to implement the tracked vehicle driving state prediction model training method or the tracked vehicle driving state prediction method of any one of the above.

[0039] According to the specific embodiments provided by the application, the following technical effects are disclosed:

[0040] The application provides a tracked vehicle driving state prediction model training method, a prediction method and products. The method comprises the following steps: obtaining measured vector data and measurable control vector data of a tracked vehicle; the measured vector data comprises longitudinal, lateral acceleration and yaw angular velocity of the tracked vehicle; the measurable control vector data comprises left and right drive motor torque, left and right driving wheel speed and transmission gear position; inputting the measurable control vector data into a trained transmission-track model to obtain actual control vector data; the actual control vector data comprises lateral and longitudinal force of the ground on the track, steering drive torque and steering resistance torque; obtaining state vector data of the tracked vehicle at the previous moment; the state vector data comprises predicted values of longitudinal, lateral speed and yaw angular velocity of the tracked vehicle; calculating prior estimation of the state vector data at the current moment by using a Kalman recursive framework to obtain prior state estimation; calculating prior estimation of the measured vector data by using the Kalman recursive framework to obtain prior measurement estimation; converting the prior state estimation, the prior measurement estimation and the actual control vector data into a difference sequence; training a recurrent neural network by using the difference sequence to obtain a tracked vehicle driving state prediction model; the output of the tracked vehicle driving state estimator is state vector prediction data; the state vector prediction data is real-time longitudinal, lateral speed and yaw angular velocity of the tracked vehicle predicted. The application combines a neural network with good nonlinear fitting characteristics into a Kalman filtering algorithm to learn potential noise characteristics, covariance matrix and Kalman gain, avoids dependence on a mechanism model and matrix inversion operation, finally realizes effective state estimation of a strong nonlinear system, and integrates Kalman filtering state estimation principles to calculate prior estimation of state quantities and measurements by using a mechanism model, and combines control quantities to perform difference sequence operation, then an internal neural network only needs to learn Kalman gain by using the difference sequence, has a clear task, and has strong learnable behavior, so that the number of parameters to be learned is small, and training efficiency is improved. The application can realize effective state estimation of a strong nonlinear system, and the number of neural network parameters to be learned is small, and training efficiency is improved. BRIEF DESCRIPTION OF DRAWINGS

[0041] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0042] Figure 1 A flowchart of a tracked vehicle driving state prediction model training method provided by an embodiment of the present application is shown in the figure.

[0043] Figure 2A schematic diagram of force analysis of a track shoe is provided for an embodiment of the present application.

[0044] Figure 3 A schematic diagram of a state estimation process based on a Kalman neural network is provided for an embodiment of the present application.

[0045] Figure 4 A schematic diagram of an internal recurrent neural network framework design of a Kalman neural network is provided for an embodiment of the present application.

[0046] Figure 5 A schematic diagram of the structure of a computer device is provided for an embodiment of the present application. DETAILED DESCRIPTION

[0047] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative work fall within the scope of protection of the present application.

[0048] The present application provides a track vehicle state prediction method based on a Kalman neural network, which comprises the following steps: considering the development of relative sliding between the track and the ground, a track force model is established; a data-driven modeling is performed on the transmission-track model to predict the ground force and steering torque on the track online; a Kalman filter based on a mechanism model is combined with a recurrent neural network to construct a Kalman neural network with a recursive structure; and the constructed network is trained to learn the Kalman gain from data to update and correct the track vehicle state. The present application is a combination of mechanism knowledge and data-driven architecture, which solves the problems of model mismatch and inability to adapt to strong nonlinearity in the application of traditional Kalman filter, and improves the interpretability of neural network, with the characteristics of fast training convergence and high state estimation accuracy.

[0049] The above-mentioned purposes, features and advantages of the present application can be more obvious and easy to understand. The present application will be described in further detail below with reference to the drawings and specific embodiments.

[0050] In an exemplary embodiment, as shown in Figure 1 a track vehicle driving state prediction model training method is provided, comprising:

[0051] S1, obtaining measurement vector data and measurable control vector data of a track vehicle; the measurement vector data comprises longitudinal, lateral acceleration and yaw rate of the track vehicle; and the measurable control vector data comprises left and right drive motor torque, left and right driven wheel speed and transmission gear position.

[0052] S2, input the measurable control vector data into the trained drive-track model to obtain actual control vector data; the actual control vector data includes: lateral and longitudinal force of the ground on the track, steering driving torque and steering resisting torque.

[0053] The training method of the drive-track model specifically comprises:

[0054] Obtain drive data of the tracked vehicle; the drive data includes: left and right drive motor torque training data, left and right driving wheel speed training data, and transmission gear training data.

[0055] Perform force analysis on the tracked vehicle to obtain vehicle force data; the vehicle force data includes: lateral and longitudinal force training data of the ground on the track of the tracked vehicle, steering driving torque training data, and steering resisting torque training data.

[0056] Take the drive data as input and the vehicle force data as output, train a neural network to obtain a drive-track model. The neural network is a long short-term memory neural network.

[0057] S3, obtain state vector data of the tracked vehicle at the previous time; the state vector data includes: predicted values of longitudinal, lateral speed and yaw angular speed of the tracked vehicle. The state vector data in this embodiment cannot be directly obtained, and each state vector data is predicted, wherein the first state vector data (i.e. the initial value) is initialized. S4, calculate the priori estimation of the state vector data at the current time by using a Kalman recursive framework to obtain the priori state estimation.

[0058] The Kalman recursive framework is a Kalman recursive framework based on a mechanism model;

[0059] The processing process of the Kalman recursive framework comprises:

[0060] A tracked vehicle dynamics model is established:

[0061]

[0062] In the formula, m is the mass of the vehicle, I z is the steering inertia of the vehicle, B is the track center distance, v x is the longitudinal speed, v y is the lateral speed, ω is the steering angular speed, i = 1, 2 respectively represent the left and right tracks, F xi is the longitudinal driving force of the ground on the two tracks, F yi is the lateral force of the ground on the two tracks, T di is the steering driving torque of the two tracks, T ri is the steering resisting torque of the two tracks, FN1 ,F N2 is the ground pressure on the left and right side tracks, g is the acceleration of gravity, f r is the ground resistance coefficient;

[0063] The model state variables are x = [v x ,v y ,ω] T , the control variables are u = [F x1 ,F x2 ,F y1 ,F y2 ,T r1 ,T r2 ,T d1 ,T d2 ] T , and the measurement variables are y = [a x ,a y ,ω] T , wherein a x ,a y are the longitudinal and lateral accelerations of the vehicle, and ω is the yaw rate;

[0064] The state transition function is:

[0065]

[0066] According to the vehicle dynamics model, the measurement function is:

[0067]

[0068] S5. Calculate the prior estimate of the measurement vector data using the Kalman recursive framework to obtain a prior measurement estimate.

[0069] S6. Convert the prior state estimate, the prior measurement estimate, and the actual control vector data into a difference sequence.

[0070] S7. Train a recurrent neural network using the difference sequence to obtain a tracked vehicle driving state prediction model; the output of the tracked vehicle driving state estimator is state vector prediction data; the state vector prediction data is the real-time longitudinal, lateral speed, and yaw rate of the tracked vehicle obtained by prediction.

[0071] The loss function of the recurrent neural network is: wherein, represents the loss function, and Δy k represents the measurement update difference.

[0072] The recurrent neural network is a long short-term memory neural network; the long short-term memory neural network comprises three independent LSTM units and a plurality of FC layers; the three independent LSTM units are respectively used for learning process noise covariance matrix, state covariance matrix and measurement covariance matrix.

[0073] The above process is realized through the following four steps in the embodiment:

[0074] Step A1, considering the development process of the relative sliding between the track and the ground, a track force model is established, first, the force analysis of each track shoe is carried out to obtain the normal force p ik , longitudinal force ε xik and lateral force ε yik of each track shoe, then the distributed force is integrated into concentrated force, i.e. the force and torque of the ground acting on the two sides of the track. Please refer to Figure 2 which describes the lateral, longitudinal sliding and ground force of each track shoe when the tracked vehicle turns, specifically: the contact force of each track shoe with the ground is decomposed into normal force p ik , longitudinal force ε xik and lateral force ε yik , wherein subscript ik represents the kth track shoe from front to back on the i side of the track.

[0075] The track force model of step A1 has the following specific content:

[0076] First, the track winding speed is calculated by the output speed of the main drive wheel of the mechatronic composite transmission system, and then the force of each track shoe is calculated according to the movement speed of each track shoe, including the normal ground pressure and the tangential horizontal force. The winding movement of the track is driven by the torque output by the main drive wheel, and the dynamics equation of the track running device is represented as follows:

[0077]

[0078] Wherein, i = 1, 2, respectively represent left and right sides, J tr is the equivalent moment of inertia of the single side track running device at the main drive wheel, is the angular acceleration of the certain side main drive wheel, T spi is the driving torque output by the main drive wheel, which is obtained by the torque of the driving motor on both sides through the front transmission mechanism, double motor coupling mechanism, transmission and side transmission mechanism of the mechatronic composite transmission system, r z is the radius of the main drive wheel, F xi is the longitudinal force of the ground acting on the certain side track, ignoring the influence of the moment of inertia and mechanical efficiency of the transmission mechanism, the transmission system model representing the relationship between the main drive torque and the driving motor torque is established as follows, wherein, i q ,i b ,ic are the front gear ratio, the two-speed transmission gear ratio, the side gear ratio, respectively, k x is the coupling mechanism planetary gear set parameters, T m1 and T m2 are the left and right side drive motor output torque, respectively.

[0079]

[0080] In the track force model, the horizontal force between the track and the ground is represented by concentrated forces (longitudinal force F xi , lateral force F yi ) and concentrated moments (steering resistance moment T ri , steering drive moment T di ). In order to more accurately simulate the dynamic behavior of the tracked vehicle and the complex contact with the ground, the force on each track shoe is calculated separately, and the ground force on each track shoe is considered as a concentrated force, which is decomposed into normal force p ik , longitudinal force ε xik and lateral force ε yik , where subscript ik represents the kth track shoe from front to back on the i side. In the case where the tracked vehicle does not significantly roll, the effect of the steering centrifugal force on the track ground pressure distribution is considered, and the track ground pressure is equivalent to a trapezoidal distribution, and the ground pressure F Ni on the two tracks is calculated as follows, where m is the mass of the tracked vehicle, g is the acceleration of gravity, h c is the height of the center of mass, B is the track center distance, ω is the yaw rate, v x is the longitudinal vehicle speed, is the lateral acceleration.

[0081]

[0082] The number of track shoes on one side of the track that contact the ground is denoted as n, and it is assumed that the track ground pressure is trapezoidal. Since the ground pressure F Ni on one side can be obtained by integrating the normal force p ik of each track shoe on the ground pressure length of the track, the calculated normal force (ground pressure) of each track shoe is shown in the following formula, where x ik is the x coordinate of the track shoe, L is the track ground length, v y is the lateral vehicle speed, is the longitudinal acceleration.

[0083]

[0084] The horizontal friction force acting on each track shoe can be calculated by the following formula, where the friction coefficient μ and the ground shear modulus K are the contact parameters of the track and the ground, and ε ikis the actual friction force, ε maxik is the maximum friction force, j ik is the slip of the track shoe, i.e. the shear of the ground at this track shoe, j ik is calculated from the slip velocity of each track shoe, p ik is the normal force of the track shoe, calculated from the above equation.

[0085]

[0086] The steering of the tracked vehicle is driven by the differential of the two driving wheels. Assuming that the elongation and contraction of the track can be neglected, the velocity of the track relative to the vehicle body is v spi = ω spi r z , where ω spi and r z are the angular velocity and radius of the driving wheel, respectively. In the vehicle coordinate system V(x,y), the slip velocity of the track shoe relative to the ground is calculated by the following equation.

[0087]

[0088] In the equation, v jxik is the longitudinal slip velocity of the track shoe relative to the ground, v jyik is the lateral slip velocity of the track shoe relative to the ground, R s is the turning radius, L is the ground contact length of the track, v spi is the velocity of the track relative to the vehicle body, and the x coordinate x ik of the track shoe is calculated by the following equation:

[0089]

[0090] In the equation, t 0ik is the time when the track shoe starts to contact the ground, t is the current time, and in the definite integral, τ is the independent variable (time) with the upper and lower limits of t and t 0ik .

[0091] According to the steering angle φ of the tracked vehicle, the slip velocities v jXik , v jYik of the track shoe are obtained by transforming to the inertial coordinate system, and then the shear j ik and the longitudinal and lateral shears j Xik , j Yik are obtained by integrating the slip velocities as follows:

[0092]

[0093] According to the relationship between friction and shear, the ground friction force acting on each track shoe is obtained. Assuming that the friction force acting on the track shoe is opposite to the direction of the sliding speed of the track shoe, the friction force on each track shoe can be decomposed into x and y directions by the following formula:

[0094]

[0095] In the formula, δ ik is the direction angle of the sliding speed of the track shoe. δ ik satisfies the following formula:

[0096]

[0097] According to the relationship between the ground shear displacement of the vehicle track, the normal force and the tangential force of each track shoe are calculated, and finally the distributed force is converted into concentrated force, so that the force and torque of the ground acting on the track are obtained:

[0098]

[0099] In the above formula, F xi and F yi are the longitudinal and transverse concentrated forces of the ground acting on the track, ε xik and ε yik are the longitudinal and transverse contact forces of the ground acting on the kth track shoe. T ri is the steering resistance torque caused by the distributed transverse contact force, and T di is the steering driving torque caused by the distributed longitudinal contact force.

[0100] After the transmission system model and the track force model are established as above, the relationship between the ground force acting on the two tracks and the steering torque and the input torque of the two drive motors under a certain transmission gear can be described.

[0101] Step A2, using neural network to data-driven modeling of transmission system model and track force model, get transmission-track model, reduce the complexity of calculation, using the model in the given two side drive motor torque input, online prediction of the track of the ground horizontal and longitudinal force and steering torque, in order to as control input to the state transition equation to calculate the prior state estimation.

[0102] The data-driven transmission-track model in step A2 has the following specific content:

[0103] A2.1, first use the transmission system model and track force model simulation to generate training data set, specific including:

[0104] In different driving conditions of a series of target vehicle speeds and target yaw angular velocities, track control of the tracked vehicle is performed using different gears, and then the input torques T m1 ,T m2 of the two side driving motors in the running data are taken as the input of the data-driven model sp1 ,ω sp2 , the transmission gear z is taken as the input of the data-driven model, and the ground lateral and longitudinal forces F xi ,F yi and the steering torque T ri ,T di calculated through the electromechanical composite transmission system and the tracked force model are taken as the output of the data-driven model, and the input and output quantities are collected to build a training database.

[0105] A2.2, then select a suitable neural network to build a data-driven transmission-track model, which specifically includes:

[0106] According to the type of input and output quantities, the type of neural network is selected. Since the driving motor torque and the ground force and torque acting on the tracked are time series data, there is a dependent relationship between the measured quantity and the historical input, so the long short-term memory neural network in the recurrent neural network is selected to predict the output quantity according to the input quantity. The long short-term memory neural network comprises an input layer, an LSTM layer and an output layer. The input layer model at any k time is:

[0107]

[0108] In the formula, T represents the input torque of the left and right driving motors, ω represents the rotational speed of the left and right driving wheels, z is the transmission gear, and each LSTM layer at any k time is composed of 64 cell units (hidden layer). The model of each cell unit is:

[0109]

[0110] In the formula, σ(·) is a sigmoid activation function; i k , is an input gate layer; f k is a forget gate layer; o k , h k is an output gate layer; c k is the cell unit state at the current time; W i , W c , W f , W o and b i , b c , b f , b oThe weight matrix and bias term to be learned of the input gate layer, the forget gate layer and the output gate layer of each cell unit are input respectively; the subscripts k and k-1 respectively represent the variable at the current moment and the variable at the last moment; each cell unit has an output h k ; the output layer adds the outputs h k of the 64 cell units to obtain the ground force and the moment of force at the current moment The superscript k represents the variable at the current moment; the output expression is as follows:

[0111]

[0112] After the data-driven transmission-track model is established, the lateral and longitudinal forces of the track, the steering driving moment and the steering resisting moment can be predicted on line under a certain transmission gear according to the torque input of the two driving motors, so as to be input into the state transition equation as control quantities to calculate the prior state estimation and the prior measurement estimation.

[0113] Step A3, combining the Kalman filter based on the mechanism model with the recurrent neural network based on the data-driven model, a Kalman neural network with a recursive structure is constructed. Please refer to Figure 3 , Figure 3 is the state estimation flowchart of the Kalman neural network described in step A3 of the present application, including an external Kalman filter state estimation framework and an internal recurrent neural network framework, which describes the principle flow of using the Kalman neural network to perform differential and recursive operations, and finally realizes state estimation.

[0114] The principle and design of the Kalman neural network of step A3 are as follows:

[0115] The principle and design of the Kalman neural network include Kalman filter theory, state estimation principle based on Kalman neural network and Kalman neural network design based on the principle.

[0116] A3.1, the Kalman filter theory analyzes the principle of Kalman filter state estimation for a nonlinear system, which is the theoretical basis for designing the internal and external recursive framework of the Kalman neural network, and specifically:

[0117] A time-varying nonlinear system is established, the system noise has a Gaussian distribution, and after discretization, the state equation is as follows:

[0118]

[0119] Where x k , y k , u k are the state value, measurement value and control quantity at k moment, respectively, x k-1Xk-1denotes the state value at time k-1 k-1 Xk-1denotes the state value at time k-1 k ,v k Q and R are the covariance matrices of process noise and measurement noise, respectively, which are Gaussian distributed with mean 0. f(·) and h(·) are the state transition function and measurement function of the nonlinear system, respectively.

[0120] The model-based Kalman filtering algorithm is a linear recursive estimation algorithm. The principle is to use the posterior state estimate at the last time step to predict the prior state estimate at the current time step and then use the measurement value y k at the current time step to correct the prior estimate to obtain the posterior state estimate at the current time step which is also called the optimal state estimate. This cycle continues until the iteration condition is met. The application of this algorithm to nonlinear systems is the extended Kalman filter, which linearizes the nonlinear problem locally through first-order Taylor expansion, improving the estimation effect of nonlinear systems. The extended Kalman filter can be described by a two-step process: state update and measurement correction, where each step calculates the estimated mean and covariance matrix. Specifically, it includes:

[0121] 1) In the state update step, the prior state estimate and the prior measurement estimate are predicted based on the posterior state estimate at the last time step and the control quantity at the current time step. The specific prediction formula is as follows:

[0122]

[0123] When f(·) and h(·) are differentiable, the extended Kalman filter linearizes them by calculating the Jacobian matrix through partial derivatives, and the values are taken at and respectively. The state transition Jacobian matrix A k and the measurement Jacobian matrix H k are calculated as follows:

[0124]

[0125] The prediction uncertainty is described by the state covariance matrix P k|k-1 and the measurement covariance matrix S k|k-1 , which are calculated as follows:

[0126]

[0127] 2) In the measurement correction step, the posterior estimate is calculated based on the prior estimate. The key to this step is to calculate the Kalman gain Kk , which is calculated by the state covariance matrix, the measurement covariance matrix and the hidden measurement matrix in the measurement covariance matrix:

[0128]

[0129] The current measurement value y k is compared with the prior measurement estimate to obtain the measurement update difference Ay k , and the Kalman gain is combined to update the posterior state estimate and the state covariance matrix P k , which is calculated as follows:

[0130]

[0131] A3.2, the state estimation principle based on Kalman neural network, specifically:

[0132] Similar to the extended Kalman filter algorithm, at each time t=k, the Kalman neural network estimates

[0133] The state update step is the same as the extended Kalman filter algorithm, which only predicts the prior estimate of the state and the measurement and Specifically, the Kalman neural network first uses the posterior state estimate at the last time k and the state transition function f(·) to calculate the current prior state estimate Then, the current prior state estimate at the last time k and the measurement function h(·) are used to calculate the current prior measurement estimate

[0134] The measurement correction step is the same as the extended Kalman filter algorithm, which uses the current measurement value y k and the prior measurement estimate to calculate the update difference Ay k , and then combines the Kalman gain K k to correct to obtain the posterior state estimate at the current time k|k-1 The difference between the extended Kalman filter algorithm and the Kalman neural network is that the Kalman gain in the Kalman neural network does not need to be calculated first covariance matrix P k|k-1 and S k|k-1 to be explicitly solved, but learned from data by recurrent neural network under unknown noise statistics.

[0135] A3.3. The principle-based Kalman neural network design of A3.2, which is designed to estimate the tracked vehicle state, includes an external Kalman recursion framework based on the mechanism model and an internal recurrent neural network framework based on data-driven. Please refer to Figure 4 , Figure 4 A3.3. The internal recurrent neural network framework design of A3.2, which describes the process of tracking each covariance matrix in the Kalman filtering principle from the data using the LSTM unit in the recurrent neural network architecture, and finally obtaining the Kalman gain.

[0136] A3.3. The external Kalman recursion framework based on the mechanism model in A3.3, specifically:

[0137] To perform the state update step described in A3.2, a tracked vehicle dynamics model is established to describe the longitudinal, lateral, and yaw motion on a horizontal ground, which is established as follows:

[0138]

[0139] where m is the vehicle mass, I z is the steering inertia of the vehicle, B is the track center distance, v x is the longitudinal velocity, v y is the lateral velocity, ω is the steering angular velocity, i = 1, 2 respectively represent the left and right tracks, F xi is the longitudinal driving force of the ground acting on the two tracks, F yi is the lateral force of the ground acting on the two tracks, T ri is the steering resistance torque of the two tracks, T di is the steering driving torque of the two tracks, F N1 ,F N2 is the ground pressure on the left and right tracks, g is the acceleration of gravity, f r is the ground resistance coefficient.

[0140] The model state variable is x = [v x , v y , ω] T , the control variable is u = [F x1 , F x2 , F y1 , F y2 , T r1 , T r2 , T d1 , T d2 ] T , and the measurement is y = [a x , a y , ω] T , where a x , ay respectively, are measured by accelerometers, and the yaw rate ω is measured by a yaw rate sensor. The control variable u is calculated online by the data-driven transmission-track model according to the torque input of the two side driving motors at a certain transmission gear.

[0141] Therefore, the state transition function is:

[0142]

[0143] According to the vehicle dynamics model, the measurement function is obtained as:

[0144]

[0145] To perform the measurement correction step described in A3.2, the state quantity, the measurement quantity, and the control quantity are processed to obtain the input quantity of the internal recurrent neural network framework so that the recurrent neural network learns the Kalman gain to correct the prior state estimation, as follows:

[0146] The Kalman filter theory described in A3.1 indicates that to obtain the Kalman gain, the measurement matrix information H k , the covariance matrix information P k|k-1 , and S k|k-1 need to be learned from the state transition information and the noise statistical characteristics.

[0147] The measurement evolution difference F1 is:

[0148] The measurement update difference F2 is:

[0149] The state evolution difference F3 is:

[0150] The state update difference F4 is:

[0151] The control quantity difference F5 is Δu k = u k - u k-1 ;

[0152] where y k-1 is the measurement value at the previous time, x is the posterior state estimation at the previous time, and u k-1 is the control quantity at the previous time. The input features are all difference sequences, F1, F3, and F5 represent the measurement evolution difference, the state evolution difference, and the control quantity difference of the state transition information, and F2 and F4 include the measurement update difference and the state update difference representing the noise statistical characteristics.

[0153] Using difference sequences as network input allows for better learning of noise characteristics, optimization of noise-induced errors, avoidance of computational errors from directly predicting state variables, and faster training convergence. To enable the recurrent neural network to fully learn the Kalman gain, a comprehensive difference sequence {F1, F2, F3, F4, F5} is selected as the input feature sequence. Specifically, in the current time step k, The Kalman neural network needs to be computed, so the available F3 and F4 features are: and

[0154] The external Kalman recursive framework processes each variable to obtain... Then, the input feature difference sequence is used for training and computation of the internal recurrent neural network, and the Kalman gain K of the recurrent neural network output is... k It is then used to correct the prior state estimate.

[0155] The design of the internal recurrent neural network framework of A3.3 is as follows:

[0156] Long Short-Term Memory (LSTM) neural networks from recurrent neural networks are chosen to learn the covariance matrix in the Kalman filter principle of the data. This recurrent neural network structure, which has long-term memory and long-term dependency, can control the flow of information through gating units, thus solving the gradient vanishing and exploding problems in traditional RNNs.

[0157] The recurrent neural network architecture uses a separate LSTM unit for tracking each layer when learning the covariance matrix. It consists of three independent LSTM units (LSTM 1, LSTM 2, and LSTM 3) and some fully connected (FC) layers. The tracking state covariance matrix P... k|k-1 First, you need to learn about process noise Q. k Based on the statistical properties, LSTM 1 was designed from... Learn and track the process noise Q, and predict the Q k Input features Δu k The data is fed into an LSTM 2 to learn and track the state covariance matrix P. k|k-1 Similarly, the P predicted by LSTM 2... k|k-1 With input feature Δy k , Δu k The measurement matrix H is input into the LSTM 3. k Calculate the measurement covariance matrix S with measurement noise R. k|k-1 Finally, combine P k|k-1 and S k|k-1 To learn the Kalman gain K at the current momentk .

[0158] The design of the dimensions of the hidden layers of each LSTM module and the input and output layers of the FC module in the recurrent neural network structure is as follows: first, the dimension of the state quantity is regarded as m, m = 3, the dimension of the measurement quantity is regarded as n, n = 3, and the dimension of the control quantity is regarded as r, r = 8; then, the dimensions of each neural network module are designed according to these dimensions.

[0159] The dimension of the hidden layer of the LSTM 1 tracking process noise Q is m 2 The dimension of the hidden layer of the LSTM 2 tracking state covariance matrix P k|k-1 is m 2 The dimension of the hidden layer of the LSTM 3 tracking measurement covariance matrix S k|k-1 is n 2 There are seven fully connected layer modules, the input layer dimension of FC 1 is the same as the hidden layer dimension m 2 of LSTM 2, and the output layer dimension is the same as the hidden layer dimension n 2 of LSTM 3; the input layer dimension of FC 2 is the sum of the hidden layer dimension m 2 of LSTM 2 and the hidden layer dimension n 2 of LSTM 3, and the output layer dimension is the same as the dimension m·n of the Kalman gain K k ; the input layer dimension of FC 3 is the sum of the hidden layer dimension n 2 of LSTM 3 and the output layer dimension m·n of FC 2, and the output layer dimension is the same as the hidden layer dimension m 2 of LSTM 2; the input layer dimension of FC 4 is the sum of the hidden layer dimension m 2 of LSTM 2 and the output layer dimension m 2 of FC 3, and the output layer dimension is the same as the hidden layer dimension m 2 of LSTM 2; the input layer dimension of FC 5 is the dimension m of the state update difference , and the output layer dimension is an integer multiple of the input layer dimension; the input layer dimension of FC 6 is the sum of the dimension m of the state evolution difference and the dimension r of the control quantity difference Δu k , and the output layer dimension is an integer multiple of the input layer dimension; the input layer dimension of FC 7 is the sum of the dimension n of the measurement evolution difference Δy k , the dimension n of the measurement update difference , and the dimension r of the control quantity difference Δu k , and the output layer dimension is an integer multiple of the input layer dimension.

[0160] Step A4, the network described in step A3 learns Kalman gain from data, gets actual control amount by using the tracked-vehicle model described in step S2, and updates the tracked-vehicle state and corrects the measurement based on the Kalman filtering principle, and finally gets the Kalman neural network-based tracked-vehicle state estimation value.

[0161] The training of the Kalman neural network-based tracked-vehicle state estimator described in step S4 includes loss function design, learning basis, and training process, and specifically includes:

[0162] The loss function of the training is designed according to unsupervised learning, so that the Kalman neural network can train itself online to learn Kalman gain and perform state estimation, and specifically includes:

[0163] Although the true value of K k is missing in the training data set, the network cannot directly calculate the loss to update the network weight. However, since there is a relationship between Δy k and K k-1 , the Kalman neural network can be trained in an end-to-end manner by calculating the loss of the measurement update difference Δy k to indirectly learn K k , using the square error loss function , and the formula is as follows:

[0164]

[0165] The basis for learning Kalman gain by using the loss function is as follows:

[0166]

[0167] Where K k-1 is the Kalman gain at time k-1, Δy k-1 is the measurement update difference at time k-1, is a self-defined difference value, and is defined as: The expression of K k-1 is as follows, where is the prior state estimation at the previous time:

[0168]

[0169] The training process of the Kalman neural network-based tracked-vehicle driving state estimator is as follows:

[0170] Step one: initialize the tracked vehicle state vector, collect the measured vector data and the measurable control vector data of the tracked vehicle in real time, the measured vector data includes: longitudinal, lateral acceleration and yaw rate of the tracked vehicle, the measurable control vector data includes: left and right drive motor torque, left and right drive wheel speed and current transmission gear selection.

[0171] Step two: input the measurable control vector data into the data-driven drive-track model of claim 3 to obtain the current ground force, steering drive torque and steering resistance torque, which are used as actual control vector data.

[0172] Step three: input the measured vector data and the actual control vector data into the Kalman neural network, after the prior estimation calculation and difference operation of the external Kalman filter framework, input the difference sequence into the internal recurrent neural network framework, the forward propagation obtains the Kalman gain, which is used by the external Kalman filter framework, the measurement correction obtains the predicted state vector, and the loss is calculated according to the loss function, then the back propagation calculation obtains the gradient value of the internal neural network parameter, finally the Adam optimizer updates the neural network parameter according to the calculated gradient value, after multiple time step update iterations to the preset loss threshold, the tracked vehicle driving state estimator based on Kalman neural network is obtained.

[0173] At present, the state estimation method based on mechanism model, such as Kalman filter and its variants, this kind of estimation method needs to establish the kinematics or dynamics model of tracked vehicle first, then combines with sensor information to estimate the state, which is suitable for linear or weak nonlinear system that can be modeled, while the tracked vehicle driving exists the dynamics coupling between each direction, the state variable coupling is obvious, the motion is strong nonlinear. The tracked force model usually uses the integral of time and position based on the modeling of ground shear displacement, which is strong nonlinear. At the same time, due to the dramatic changes and random disturbances of the tracked vehicle driving ground conditions, the mismatch between the nominal model and the actual controlled object will occur. Strong nonlinearity and model mismatch lead to poor effect of state estimation method based on mechanism model. The data-driven state estimation method does not need to establish a complex high-precision model, but learns the relationship between system input and output through data, which has good applicability, but lacks interpretability, and the black box characteristics will lead to many parameters to be learned and long training time. In order to overcome the limitations of the above two types of state estimation methods, combining mechanism knowledge and data-driven state estimation method is a trend.

[0174] The application combines a neural network with good nonlinear fitting characteristics into a Kalman filtering algorithm to learn potential noise characteristics, covariance matrices and Kalman gains, avoids dependence on a mechanism model and matrix inversion operations, and finally realizes effective state estimation of a strong nonlinear system. The application integrates the Kalman filtering state estimation principle, first calculates the priori estimation of state quantities and measurements by using a mechanism model, and combines the control quantity to perform a difference sequence operation, then the internal neural network only needs to learn the Kalman gain by using the difference sequence, has a clear task, and the neural network learns the process noise covariance matrix, the state covariance matrix and the measurement covariance matrix by designing three LSTM modules, has strong learning behavior explainability, makes the to-be-learned parameters less, and improves the training efficiency.

[0175] In another exemplary embodiment, a tracked vehicle driving state prediction method is provided, and the method comprises:

[0176] obtaining state vector data of a previous moment, real-time measurement vector data and real-time measurable control vector data of a tracked vehicle to be predicted;

[0177] inputting the state vector data of the previous moment, the real-time measurement vector data and the real-time measurable control vector data into a tracked vehicle driving state prediction model to obtain real-time longitudinal, lateral speed and yaw angular velocity of the tracked vehicle to be predicted; the tracked vehicle driving state prediction model is trained by the tracked vehicle driving state prediction model training method.

[0178] In an exemplary embodiment, a computer device is provided, which can be a server or a terminal, and an internal structure diagram thereof can be as shown in Figure 5 The computer device comprises a processor, a memory, an input / output interface (I / O) and a communication interface. The processor, the memory and the input / output interface are connected through a system bus, and the communication interface is connected to the system bus through the input / output interface. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device comprises a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operating system and the computer program in the non-volatile storage medium to run. The input / output interface of the computer device is used to exchange information between the processor and external devices. The communication interface of the computer device is used to communicate with external terminals through network connection. The computer program is executed by the processor to implement a tracked vehicle driving state prediction model training method or a tracked vehicle driving state prediction method.

[0179] Those skilled in the art can understand that, Figure 5The structure shown in the figure is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0180] In an exemplary embodiment, a computer device is also provided, including a memory and a processor, the memory storing a computer program, and the processor implementing the steps in the above method embodiments when executing the computer program.

[0181] In an exemplary embodiment, a computer readable storage medium is provided, storing a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0182] In an exemplary embodiment, a computer program product is provided, including a computer program, and the computer program implements the steps in the above method embodiments when executed by a processor.

[0183] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0184] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a non-volatile computer readable storage medium, and when the computer program is executed, the processes of the above-mentioned embodiments of the methods can be included. Any reference to memory, database or other medium used in the embodiments provided in the present application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical storage, high-density embedded non-volatile memory, resistive memory (ReRAM), magnetoresistive random access memory (MRAM), ferroelectric memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. As an illustration but not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0185] The database involved in the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, etc., without being limited thereto. The processor involved in the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, etc., without being limited thereto.

[0186] The technical features of the above embodiments can be combined arbitrarily. In order to make the description concise, all possible combinations of the technical features in the above embodiments are not described, however, as long as the combinations of the technical features do not exist contradictory, they should be considered as the scope of the present application.

[0187] The principles and implementation modes of the present application are described by applying specific examples in the present application. The above-mentioned embodiments are only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, the specific implementation mode and application range will be changed. In conclusion, the content of the present application should not be understood as a limitation.

Claims

1. A training method for a tracked vehicle driving state prediction model, characterized in that, The training method for the tracked vehicle driving state prediction model includes: Acquire measurement vector data and measurable control vector data of the tracked vehicle; the measurement vector data includes: longitudinal and lateral acceleration and yaw rate of the tracked vehicle; the measurable control vector data includes: torque of the left and right drive motors, speed of the left and right drive wheels and gear position of the transmission; The measurable control vector data is input into the trained transmission-track model to obtain the actual control vector data; the actual control vector data includes: the lateral and longitudinal forces on the track from the ground, the steering driving torque, and the steering resistance torque; Obtain the state vector data of the tracked vehicle at the previous moment; the state vector data includes: predicted values ​​of the longitudinal and lateral velocities and yaw rate of the tracked vehicle. The prior estimate of the state vector data at the current moment is calculated using the Kalman recursive framework to obtain the prior state estimate; The prior estimate of the measurement vector data is calculated using the Kalman recursive framework to obtain the prior measurement estimate; The prior state estimate, the prior measurement estimate, and the actual control vector data are converted into difference sequences. A recurrent neural network is trained using the differential sequence to obtain a tracked vehicle driving state prediction model; the output of the tracked vehicle driving state prediction model is state vector prediction data; the state vector prediction data is the predicted real-time longitudinal, lateral, and yaw rates of the tracked vehicle. The training method for the drive-track model specifically includes: Acquire drive data for tracked vehicles; the drive data includes: left and right drive motor torque training data, left and right drive wheel speed training data, and transmission gear training data; Force analysis is performed on the tracked vehicle to obtain vehicle force data; the vehicle force data includes: training data of lateral and longitudinal forces on the track of the tracked vehicle, training data of steering drive torque and training data of steering resistance torque. Using the driving data as input and the vehicle force data as output, a neural network is trained to obtain a transmission-track model; The Kalman recursive framework is a Kalman recursive framework based on a mechanistic model; The processing steps of the Kalman recursive framework include: Establish a dynamic model for tracked vehicles: In the formula, For longitudinal acceleration, For lateral acceleration, Let I be the yaw acceleration, i = 1, 2 represent the left and right tracks respectively, m be the vehicle mass, and I be the yaw acceleration. z V is the vehicle's steering inertia, B is the track center distance, and v is the center distance of the tracks. x Let v be the longitudinal velocity. y ω is the lateral velocity, ω is the angular velocity of turning, and F is the lateral velocity. xi F is the longitudinal driving force exerted by the ground on both sides of the tracks. yi T is the lateral force exerted by the ground on both sides of the tracks. di T is the steering driving torque exerted on both tracks. ri F is the steering resistance torque experienced by both tracks. N1 ,F N2 The ground pressure on the left and right tracks is given by f, g is the acceleration due to gravity, and f is the ground pressure on the left and right tracks. r This refers to the ground drag coefficient; The model's state variables are x = [v x ,v y ,ω] T The control quantity is u = [F x1 ,F x2 ,F y1 ,F y2 ,T r1 ,T r2 ,T d1 ,T d2 ] T The measurement is y = [a x ,a y ,ω] T , where a x ,a y These are the longitudinal and lateral accelerations of the vehicle, respectively, and ω is the yaw rate. The state transition function is: Based on the vehicle dynamics model, the measurement function is obtained as follows:

2. The training method for predicting the driving state of tracked vehicles according to claim 1, characterized in that, The neural network is a long short-term memory neural network.

3. The training method for predicting the driving state of tracked vehicles according to claim 1, characterized in that, The loss function of the recurrent neural network is: in, Let Δy represent the loss function. k This indicates the measurement update difference.

4. The training method for predicting the driving state of tracked vehicles according to claim 1, characterized in that, The recurrent neural network includes: three independent LSTM units and several FC layers; Three independent LSTM units are used to learn the process noise covariance matrix, the state covariance matrix, and the measurement covariance matrix, respectively.

5. A method for predicting the driving state of a tracked vehicle, characterized in that, The method for predicting the driving status of tracked vehicles includes: Acquire the previous state vector data, real-time measurement vector data, and real-time measurable control vector data of the tracked vehicle to be predicted; The state vector data of the previous moment, the real-time measurement vector data, and the real-time measurable control vector data are input into the tracked vehicle driving state prediction model to obtain the real-time longitudinal, lateral, and yaw rates of the tracked vehicle to be predicted; the tracked vehicle driving state prediction model is trained by the tracked vehicle driving state prediction model training method according to any one of claims 1-4.

6. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the tracked vehicle driving state prediction model training method according to any one of claims 1-4 or the tracked vehicle driving state prediction method according to claim 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the training method for predicting the driving state of a tracked vehicle as described in any one of claims 1-4 or the method for predicting the driving state of a tracked vehicle as described in claim 5.

8. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the training method for predicting the driving state of a tracked vehicle as described in any one of claims 1-4 or the method for predicting the driving state of a tracked vehicle as described in claim 5.

Images